OpenAI Exposes Russian, Iranian Ops Planting Fake News

OpenAI has shut down ChatGPT accounts tied to two state-linked influence operations, one Russian and one Iranian. Neither relied mainly on social media. Both used invented identities to get their material into legitimate media outlets, according to the company's latest investigation report.

That is a change of approach. OpenAI had already exposed a separate Russian campaign that ran on social media and a fake think tank. The two new cases show propagandists moving away from posting on platforms and toward placing content in outlets that readers already trust.

Dark Clark: a Russian campaign aimed at Latin America

The Russian operation, which OpenAI calls "Dark Clark," focused on Latin America. It had two goals: discredit Ukraine and unsettle local politics in the region.

The method had several layers:

1. A captured think tank. The operators ran a think tank through a fictitious persona. That gave their material an institutional face.
2. Unwitting local staff. According to OpenAI, the actors probably recruited local staffers who did not know who they were really working for.
3. Fabricated evidence. The campaign spread fake audio files and forged documents.

The fakes did not go unnoticed. In Ecuador and Peru, they led to fact-checks and official denials. For an influence operation, that kind of reaction counts as reach. Authorities and journalists had to spend time responding to invented material.

OpenAI places Dark Clark at category 5 out of 6 on the Breakout Scale, which it uses to rate how far an operation spreads beyond its starting point. The deciding factor was that politicians reacted to the content. The company says this is the first category 5 case in two and a half years of threat reporting. That makes Dark Clark the most effective operation OpenAI has documented by its own measure.

Bogus Bylines: Iran's fake reporters

The Iranian operation, "Bogus Bylines," worked like a fake newsroom. Its operators created seven fictional journalists and used them to place nearly 100 articles about the US-Iran conflict in online outlets around the world.

The actors also tried social media. They generated comments on platforms, but those got almost no traction. The pattern is telling: the bylined articles found a home, while the social media activity went largely unseen.

What the AI actually did

The report's most useful detail may be how both groups used ChatGPT. OpenAI says the models were not mainly writing headline-grabbing fakes. They were used for two less visible jobs:

• Internal reporting. The operators used AI to produce reports about their own work, much like staff writing updates for their managers.
• Translation and adaptation. AI helped adjust propaganda for different languages, a practical need for campaigns aimed at several countries at once.

This fits a wider view of how AI shows up in influence work. It acts less like a factory for viral deception and more like an efficiency tool for the routine admin and localization that any multilingual campaign needs.

Why the media route matters

Both campaigns chose the same path: get into real outlets instead of building a following on social media. There is a clear logic to that. A story under a credible-looking byline, or a report from an apparently established think tank, comes with trust built in. Fake accounts have to earn that trust, and platform moderation often removes them first.

The cost is that this approach needs more groundwork. Dark Clark needed a persona to run a think tank and local people to staff it. Bogus Bylines needed seven believable journalist identities that editors would accept. These are slow, deliberate operations, not mass posting.

The Bigger Picture

For CyrioX readers, the main lesson is that the risk is shifting from the platform to the publication. Outlets that accept freelance or contributed pieces, and think tanks that take on local partners, now look like real entry points for state-backed actors. Checking who a writer actually is may matter as much as checking the facts in the story.

The case also puts tools like ChatGPT text watermarks and Google's public SynthID detector in context. Provenance signals may help, but the operators here used AI mostly for translation and internal paperwork. Content that has been only partly machine-assisted may not stand out much. The fabricated audio in Dark Clark is a reminder that convincing synthetic media can reach politicians before fact-checkers catch up.

It is worth watching whether other AI providers report similar placement tactics, and whether more cases reach category 5. If they do, the Dark Clark rating may turn out to be an early example of a broader trend rather than a one-off.